block/blk-mq.c

   1 /*
   2  * Block multiqueue core code
   3  *
   4  * Copyright (C) 2013-2014 Jens Axboe
   5  * Copyright (C) 2013-2014 Christoph Hellwig
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/mm.h>
  13 #include <linux/init.h>
  14 #include <linux/slab.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/smp.h>
  17 #include <linux/llist.h>
  18 #include <linux/list_sort.h>
  19 #include <linux/cpu.h>
  20 #include <linux/cache.h>
  21 #include <linux/sched/sysctl.h>
  22 #include <linux/delay.h>
  23
  24 #include <trace/events/block.h>
  25
  26 #include <linux/blk-mq.h>
  27 #include "blk.h"
  28 #include "blk-mq.h"
  29 #include "blk-mq-tag.h"
  30
  31 static DEFINE_MUTEX(all_q_mutex);
  32 static LIST_HEAD(all_q_list);
  33
  34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  35
  36 /*
  37  * Check if any of the ctx's have pending work in this hardware queue
  38  */
  39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  40 {
  41         unsigned int i;
  42
  43         for (i = 0; i < hctx->ctx_map.map_size; i++)
  44                 if (hctx->ctx_map.map[i].word)
  45                         return true;
  46
  47         return false;
  48 }
  49
  50 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  51                                               struct blk_mq_ctx *ctx)
  52 {
  53         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  54 }
  55
  56 #define CTX_TO_BIT(hctx, ctx)   \
  57         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  58
  59 /*
  60  * Mark this ctx as having pending work in this hardware queue
  61  */
  62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  63                                      struct blk_mq_ctx *ctx)
  64 {
  65         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  66
  67         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  68                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  69 }
  70
  71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  72                                       struct blk_mq_ctx *ctx)
  73 {
  74         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  75
  76         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  77 }
  78
  79 static int blk_mq_queue_enter(struct request_queue *q)
  80 {
  81         while (true) {
  82                 int ret;
  83
  84                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  85                         return 0;
  86
  87                 ret = wait_event_interruptible(q->mq_freeze_wq,
  88                                 !q->mq_freeze_depth || blk_queue_dying(q));
  89                 if (blk_queue_dying(q))
  90                         return -ENODEV;
  91                 if (ret)
  92                         return ret;
  93         }
  94 }
  95
  96 static void blk_mq_queue_exit(struct request_queue *q)
  97 {
  98         percpu_ref_put(&q->mq_usage_counter);
  99 }
 100
 101 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 102 {
 103         struct request_queue *q =
 104                 container_of(ref, struct request_queue, mq_usage_counter);
 105
 106         wake_up_all(&q->mq_freeze_wq);
 107 }
 108
 109 /*
 110  * Guarantee no request is in use, so we can change any data structure of
 111  * the queue afterward.
 112  */
 113 void blk_mq_freeze_queue(struct request_queue *q)
 114 {
 115         bool freeze;
 116
 117         spin_lock_irq(q->queue_lock);
 118         freeze = !q->mq_freeze_depth++;
 119         spin_unlock_irq(q->queue_lock);
 120
 121         if (freeze) {
 122                 percpu_ref_kill(&q->mq_usage_counter);
 123                 blk_mq_run_queues(q, false);
 124         }
 125         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 126 }
 127
 128 static void blk_mq_unfreeze_queue(struct request_queue *q)
 129 {
 130         bool wake;
 131
 132         spin_lock_irq(q->queue_lock);
 133         wake = !--q->mq_freeze_depth;
 134         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 135         spin_unlock_irq(q->queue_lock);
 136         if (wake) {
 137                 percpu_ref_reinit(&q->mq_usage_counter);
 138                 wake_up_all(&q->mq_freeze_wq);
 139         }
 140 }
 141
 142 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 143 {
 144         return blk_mq_has_free_tags(hctx->tags);
 145 }
 146 EXPORT_SYMBOL(blk_mq_can_queue);
 147
 148 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 149                                struct request *rq, unsigned int rw_flags)
 150 {
 151         if (blk_queue_io_stat(q))
 152                 rw_flags |= REQ_IO_STAT;
 153
 154         INIT_LIST_HEAD(&rq->queuelist);
 155         /* csd/requeue_work/fifo_time is initialized before use */
 156         rq->q = q;
 157         rq->mq_ctx = ctx;
 158         rq->cmd_flags |= rw_flags;
 159         /* do not touch atomic flags, it needs atomic ops against the timer */
 160         rq->cpu = -1;
 161         INIT_HLIST_NODE(&rq->hash);
 162         RB_CLEAR_NODE(&rq->rb_node);
 163         rq->rq_disk = NULL;
 164         rq->part = NULL;
 165         rq->start_time = jiffies;
 166 #ifdef CONFIG_BLK_CGROUP
 167         rq->rl = NULL;
 168         set_start_time_ns(rq);
 169         rq->io_start_time_ns = 0;
 170 #endif
 171         rq->nr_phys_segments = 0;
 172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 173         rq->nr_integrity_segments = 0;
 174 #endif
 175         rq->special = NULL;
 176         /* tag was already set */
 177         rq->errors = 0;
 178
 179         rq->cmd = rq->__cmd;
 180
 181         rq->extra_len = 0;
 182         rq->sense_len = 0;
 183         rq->resid_len = 0;
 184         rq->sense = NULL;
 185
 186         INIT_LIST_HEAD(&rq->timeout_list);
 187         rq->timeout = 0;
 188
 189         rq->end_io = NULL;
 190         rq->end_io_data = NULL;
 191         rq->next_rq = NULL;
 192
 193         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 194 }
 195
 196 static struct request *
 197 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 198 {
 199         struct request *rq;
 200         unsigned int tag;
 201
 202         tag = blk_mq_get_tag(data);
 203         if (tag != BLK_MQ_TAG_FAIL) {
 204                 rq = data->hctx->tags->rqs[tag];
 205
 206                 if (blk_mq_tag_busy(data->hctx)) {
 207                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 208                         atomic_inc(&data->hctx->nr_active);
 209                 }
 210
 211                 rq->tag = tag;
 212                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 213                 return rq;
 214         }
 215
 216         return NULL;
 217 }
 218
 219 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 220                 bool reserved)
 221 {
 222         struct blk_mq_ctx *ctx;
 223         struct blk_mq_hw_ctx *hctx;
 224         struct request *rq;
 225         struct blk_mq_alloc_data alloc_data;
 226         int ret;
 227
 228         ret = blk_mq_queue_enter(q);
 229         if (ret)
 230                 return ERR_PTR(ret);
 231
 232         ctx = blk_mq_get_ctx(q);
 233         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 234         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 235                         reserved, ctx, hctx);
 236
 237         rq = __blk_mq_alloc_request(&alloc_data, rw);
 238         if (!rq && (gfp & __GFP_WAIT)) {
 239                 __blk_mq_run_hw_queue(hctx);
 240                 blk_mq_put_ctx(ctx);
 241
 242                 ctx = blk_mq_get_ctx(q);
 243                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 244                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 245                                 hctx);
 246                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 247                 ctx = alloc_data.ctx;
 248         }
 249         blk_mq_put_ctx(ctx);
 250         if (!rq)
 251                 return ERR_PTR(-EWOULDBLOCK);
 252         return rq;
 253 }
 254 EXPORT_SYMBOL(blk_mq_alloc_request);
 255
 256 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 257                                   struct blk_mq_ctx *ctx, struct request *rq)
 258 {
 259         const int tag = rq->tag;
 260         struct request_queue *q = rq->q;
 261
 262         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 263                 atomic_dec(&hctx->nr_active);
 264         rq->cmd_flags = 0;
 265
 266         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 267         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 268         blk_mq_queue_exit(q);
 269 }
 270
 271 void blk_mq_free_request(struct request *rq)
 272 {
 273         struct blk_mq_ctx *ctx = rq->mq_ctx;
 274         struct blk_mq_hw_ctx *hctx;
 275         struct request_queue *q = rq->q;
 276
 277         ctx->rq_completed[rq_is_sync(rq)]++;
 278
 279         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 280         __blk_mq_free_request(hctx, ctx, rq);
 281 }
 282
 283 /*
 284  * Clone all relevant state from a request that has been put on hold in
 285  * the flush state machine into the preallocated flush request that hangs
 286  * off the request queue.
 287  *
 288  * For a driver the flush request should be invisible, that's why we are
 289  * impersonating the original request here.
 290  */
 291 void blk_mq_clone_flush_request(struct request *flush_rq,
 292                 struct request *orig_rq)
 293 {
 294         struct blk_mq_hw_ctx *hctx =
 295                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 296
 297         flush_rq->mq_ctx = orig_rq->mq_ctx;
 298         flush_rq->tag = orig_rq->tag;
 299         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 300                 hctx->cmd_size);
 301 }
 302
 303 inline void __blk_mq_end_request(struct request *rq, int error)
 304 {
 305         blk_account_io_done(rq);
 306
 307         if (rq->end_io) {
 308                 rq->end_io(rq, error);
 309         } else {
 310                 if (unlikely(blk_bidi_rq(rq)))
 311                         blk_mq_free_request(rq->next_rq);
 312                 blk_mq_free_request(rq);
 313         }
 314 }
 315 EXPORT_SYMBOL(__blk_mq_end_request);
 316
 317 void blk_mq_end_request(struct request *rq, int error)
 318 {
 319         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 320                 BUG();
 321         __blk_mq_end_request(rq, error);
 322 }
 323 EXPORT_SYMBOL(blk_mq_end_request);
 324
 325 static void __blk_mq_complete_request_remote(void *data)
 326 {
 327         struct request *rq = data;
 328
 329         rq->q->softirq_done_fn(rq);
 330 }
 331
 332 static void blk_mq_ipi_complete_request(struct request *rq)
 333 {
 334         struct blk_mq_ctx *ctx = rq->mq_ctx;
 335         bool shared = false;
 336         int cpu;
 337
 338         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 339                 rq->q->softirq_done_fn(rq);
 340                 return;
 341         }
 342
 343         cpu = get_cpu();
 344         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 345                 shared = cpus_share_cache(cpu, ctx->cpu);
 346
 347         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 348                 rq->csd.func = __blk_mq_complete_request_remote;
 349                 rq->csd.info = rq;
 350                 rq->csd.flags = 0;
 351                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 352         } else {
 353                 rq->q->softirq_done_fn(rq);
 354         }
 355         put_cpu();
 356 }
 357
 358 void __blk_mq_complete_request(struct request *rq)
 359 {
 360         struct request_queue *q = rq->q;
 361
 362         if (!q->softirq_done_fn)
 363                 blk_mq_end_request(rq, rq->errors);
 364         else
 365                 blk_mq_ipi_complete_request(rq);
 366 }
 367
 368 /**
 369  * blk_mq_complete_request - end I/O on a request
 370  * @rq:         the request being processed
 371  *
 372  * Description:
 373  *      Ends all I/O on a request. It does not handle partial completions.
 374  *      The actual completion happens out-of-order, through a IPI handler.
 375  **/
 376 void blk_mq_complete_request(struct request *rq)
 377 {
 378         struct request_queue *q = rq->q;
 379
 380         if (unlikely(blk_should_fake_timeout(q)))
 381                 return;
 382         if (!blk_mark_rq_complete(rq))
 383                 __blk_mq_complete_request(rq);
 384 }
 385 EXPORT_SYMBOL(blk_mq_complete_request);
 386
 387 void blk_mq_start_request(struct request *rq)
 388 {
 389         struct request_queue *q = rq->q;
 390
 391         trace_block_rq_issue(q, rq);
 392
 393         rq->resid_len = blk_rq_bytes(rq);
 394         if (unlikely(blk_bidi_rq(rq)))
 395                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 396
 397         blk_add_timer(rq);
 398
 399         /*
 400          * Ensure that ->deadline is visible before set the started
 401          * flag and clear the completed flag.
 402          */
 403         smp_mb__before_atomic();
 404
 405         /*
 406          * Mark us as started and clear complete. Complete might have been
 407          * set if requeue raced with timeout, which then marked it as
 408          * complete. So be sure to clear complete again when we start
 409          * the request, otherwise we'll ignore the completion event.
 410          */
 411         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 412                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 413         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 414                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 415
 416         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 417                 /*
 418                  * Make sure space for the drain appears.  We know we can do
 419                  * this because max_hw_segments has been adjusted to be one
 420                  * fewer than the device can handle.
 421                  */
 422                 rq->nr_phys_segments++;
 423         }
 424 }
 425 EXPORT_SYMBOL(blk_mq_start_request);
 426
 427 static void __blk_mq_requeue_request(struct request *rq)
 428 {
 429         struct request_queue *q = rq->q;
 430
 431         trace_block_rq_requeue(q, rq);
 432
 433         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 434                 if (q->dma_drain_size && blk_rq_bytes(rq))
 435                         rq->nr_phys_segments--;
 436         }
 437 }
 438
 439 void blk_mq_requeue_request(struct request *rq)
 440 {
 441         __blk_mq_requeue_request(rq);
 442         blk_clear_rq_complete(rq);
 443
 444         BUG_ON(blk_queued_rq(rq));
 445         blk_mq_add_to_requeue_list(rq, true);
 446 }
 447 EXPORT_SYMBOL(blk_mq_requeue_request);
 448
 449 static void blk_mq_requeue_work(struct work_struct *work)
 450 {
 451         struct request_queue *q =
 452                 container_of(work, struct request_queue, requeue_work);
 453         LIST_HEAD(rq_list);
 454         struct request *rq, *next;
 455         unsigned long flags;
 456
 457         spin_lock_irqsave(&q->requeue_lock, flags);
 458         list_splice_init(&q->requeue_list, &rq_list);
 459         spin_unlock_irqrestore(&q->requeue_lock, flags);
 460
 461         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 462                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 463                         continue;
 464
 465                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 466                 list_del_init(&rq->queuelist);
 467                 blk_mq_insert_request(rq, true, false, false);
 468         }
 469
 470         while (!list_empty(&rq_list)) {
 471                 rq = list_entry(rq_list.next, struct request, queuelist);
 472                 list_del_init(&rq->queuelist);
 473                 blk_mq_insert_request(rq, false, false, false);
 474         }
 475
 476         /*
 477          * Use the start variant of queue running here, so that running
 478          * the requeue work will kick stopped queues.
 479          */
 480         blk_mq_start_hw_queues(q);
 481 }
 482
 483 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 484 {
 485         struct request_queue *q = rq->q;
 486         unsigned long flags;
 487
 488         /*
 489          * We abuse this flag that is otherwise used by the I/O scheduler to
 490          * request head insertation from the workqueue.
 491          */
 492         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 493
 494         spin_lock_irqsave(&q->requeue_lock, flags);
 495         if (at_head) {
 496                 rq->cmd_flags |= REQ_SOFTBARRIER;
 497                 list_add(&rq->queuelist, &q->requeue_list);
 498         } else {
 499                 list_add_tail(&rq->queuelist, &q->requeue_list);
 500         }
 501         spin_unlock_irqrestore(&q->requeue_lock, flags);
 502 }
 503 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 504
 505 void blk_mq_kick_requeue_list(struct request_queue *q)
 506 {
 507         kblockd_schedule_work(&q->requeue_work);
 508 }
 509 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 510
 511 static inline bool is_flush_request(struct request *rq, unsigned int tag)
 512 {
 513         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 514                         rq->q->flush_rq->tag == tag);
 515 }
 516
 517 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 518 {
 519         struct request *rq = tags->rqs[tag];
 520
 521         if (!is_flush_request(rq, tag))
 522                 return rq;
 523
 524         return rq->q->flush_rq;
 525 }
 526 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 527
 528 struct blk_mq_timeout_data {
 529         unsigned long next;
 530         unsigned int next_set;
 531 };
 532
 533 static void blk_mq_rq_timed_out(struct request *req)
 534 {
 535         struct blk_mq_ops *ops = req->q->mq_ops;
 536         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 537
 538         /*
 539          * We know that complete is set at this point. If STARTED isn't set
 540          * anymore, then the request isn't active and the "timeout" should
 541          * just be ignored. This can happen due to the bitflag ordering.
 542          * Timeout first checks if STARTED is set, and if it is, assumes
 543          * the request is active. But if we race with completion, then
 544          * we both flags will get cleared. So check here again, and ignore
 545          * a timeout event with a request that isn't active.
 546          */
 547         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 548                 return;
 549
 550         if (ops->timeout)
 551                 ret = ops->timeout(req);
 552
 553         switch (ret) {
 554         case BLK_EH_HANDLED:
 555                 __blk_mq_complete_request(req);
 556                 break;
 557         case BLK_EH_RESET_TIMER:
 558                 blk_add_timer(req);
 559                 blk_clear_rq_complete(req);
 560                 break;
 561         case BLK_EH_NOT_HANDLED:
 562                 break;
 563         default:
 564                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 565                 break;
 566         }
 567 }
 568
 569 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 570                 struct request *rq, void *priv, bool reserved)
 571 {
 572         struct blk_mq_timeout_data *data = priv;
 573
 574         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 575                 return;
 576
 577         if (time_after_eq(jiffies, rq->deadline)) {
 578                 if (!blk_mark_rq_complete(rq))
 579                         blk_mq_rq_timed_out(rq);
 580         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 581                 data->next = rq->deadline;
 582                 data->next_set = 1;
 583         }
 584 }
 585
 586 static void blk_mq_rq_timer(unsigned long priv)
 587 {
 588         struct request_queue *q = (struct request_queue *)priv;
 589         struct blk_mq_timeout_data data = {
 590                 .next           = 0,
 591                 .next_set       = 0,
 592         };
 593         struct blk_mq_hw_ctx *hctx;
 594         int i;
 595
 596         queue_for_each_hw_ctx(q, hctx, i) {
 597                 /*
 598                  * If not software queues are currently mapped to this
 599                  * hardware queue, there's nothing to check
 600                  */
 601                 if (!hctx->nr_ctx || !hctx->tags)
 602                         continue;
 603
 604                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 605         }
 606
 607         if (data.next_set) {
 608                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 609                 mod_timer(&q->timeout, data.next);
 610         } else {
 611                 queue_for_each_hw_ctx(q, hctx, i)
 612                         blk_mq_tag_idle(hctx);
 613         }
 614 }
 615
 616 /*
 617  * Reverse check our software queue for entries that we could potentially
 618  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 619  * too much time checking for merges.
 620  */
 621 static bool blk_mq_attempt_merge(struct request_queue *q,
 622                                  struct blk_mq_ctx *ctx, struct bio *bio)
 623 {
 624         struct request *rq;
 625         int checked = 8;
 626
 627         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 628                 int el_ret;
 629
 630                 if (!checked--)
 631                         break;
 632
 633                 if (!blk_rq_merge_ok(rq, bio))
 634                         continue;
 635
 636                 el_ret = blk_try_merge(rq, bio);
 637                 if (el_ret == ELEVATOR_BACK_MERGE) {
 638                         if (bio_attempt_back_merge(q, rq, bio)) {
 639                                 ctx->rq_merged++;
 640                                 return true;
 641                         }
 642                         break;
 643                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 644                         if (bio_attempt_front_merge(q, rq, bio)) {
 645                                 ctx->rq_merged++;
 646                                 return true;
 647                         }
 648                         break;
 649                 }
 650         }
 651
 652         return false;
 653 }
 654
 655 /*
 656  * Process software queues that have been marked busy, splicing them
 657  * to the for-dispatch
 658  */
 659 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 660 {
 661         struct blk_mq_ctx *ctx;
 662         int i;
 663
 664         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 665                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 666                 unsigned int off, bit;
 667
 668                 if (!bm->word)
 669                         continue;
 670
 671                 bit = 0;
 672                 off = i * hctx->ctx_map.bits_per_word;
 673                 do {
 674                         bit = find_next_bit(&bm->word, bm->depth, bit);
 675                         if (bit >= bm->depth)
 676                                 break;
 677
 678                         ctx = hctx->ctxs[bit + off];
 679                         clear_bit(bit, &bm->word);
 680                         spin_lock(&ctx->lock);
 681                         list_splice_tail_init(&ctx->rq_list, list);
 682                         spin_unlock(&ctx->lock);
 683
 684                         bit++;
 685                 } while (1);
 686         }
 687 }
 688
 689 /*
 690  * Run this hardware queue, pulling any software queues mapped to it in.
 691  * Note that this function currently has various problems around ordering
 692  * of IO. In particular, we'd like FIFO behaviour on handling existing
 693  * items on the hctx->dispatch list. Ignore that for now.
 694  */
 695 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 696 {
 697         struct request_queue *q = hctx->queue;
 698         struct request *rq;
 699         LIST_HEAD(rq_list);
 700         int queued;
 701
 702         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 703
 704         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 705                 return;
 706
 707         hctx->run++;
 708
 709         /*
 710          * Touch any software queue that has pending entries.
 711          */
 712         flush_busy_ctxs(hctx, &rq_list);
 713
 714         /*
 715          * If we have previous entries on our dispatch list, grab them
 716          * and stuff them at the front for more fair dispatch.
 717          */
 718         if (!list_empty_careful(&hctx->dispatch)) {
 719                 spin_lock(&hctx->lock);
 720                 if (!list_empty(&hctx->dispatch))
 721                         list_splice_init(&hctx->dispatch, &rq_list);
 722                 spin_unlock(&hctx->lock);
 723         }
 724
 725         /*
 726          * Now process all the entries, sending them to the driver.
 727          */
 728         queued = 0;
 729         while (!list_empty(&rq_list)) {
 730                 int ret;
 731
 732                 rq = list_first_entry(&rq_list, struct request, queuelist);
 733                 list_del_init(&rq->queuelist);
 734
 735                 ret = q->mq_ops->queue_rq(hctx, rq, list_empty(&rq_list));
 736                 switch (ret) {
 737                 case BLK_MQ_RQ_QUEUE_OK:
 738                         queued++;
 739                         continue;
 740                 case BLK_MQ_RQ_QUEUE_BUSY:
 741                         list_add(&rq->queuelist, &rq_list);
 742                         __blk_mq_requeue_request(rq);
 743                         break;
 744                 default:
 745                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 746                 case BLK_MQ_RQ_QUEUE_ERROR:
 747                         rq->errors = -EIO;
 748                         blk_mq_end_request(rq, rq->errors);
 749                         break;
 750                 }
 751
 752                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 753                         break;
 754         }
 755
 756         if (!queued)
 757                 hctx->dispatched[0]++;
 758         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 759                 hctx->dispatched[ilog2(queued) + 1]++;
 760
 761         /*
 762          * Any items that need requeuing? Stuff them into hctx->dispatch,
 763          * that is where we will continue on next queue run.
 764          */
 765         if (!list_empty(&rq_list)) {
 766                 spin_lock(&hctx->lock);
 767                 list_splice(&rq_list, &hctx->dispatch);
 768                 spin_unlock(&hctx->lock);
 769         }
 770 }
 771
 772 /*
 773  * It'd be great if the workqueue API had a way to pass
 774  * in a mask and had some smarts for more clever placement.
 775  * For now we just round-robin here, switching for every
 776  * BLK_MQ_CPU_WORK_BATCH queued items.
 777  */
 778 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 779 {
 780         int cpu = hctx->next_cpu;
 781
 782         if (--hctx->next_cpu_batch <= 0) {
 783                 int next_cpu;
 784
 785                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 786                 if (next_cpu >= nr_cpu_ids)
 787                         next_cpu = cpumask_first(hctx->cpumask);
 788
 789                 hctx->next_cpu = next_cpu;
 790                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 791         }
 792
 793         return cpu;
 794 }
 795
 796 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 797 {
 798         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 799                 return;
 800
 801         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 802                 __blk_mq_run_hw_queue(hctx);
 803         else if (hctx->queue->nr_hw_queues == 1)
 804                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 805         else {
 806                 unsigned int cpu;
 807
 808                 cpu = blk_mq_hctx_next_cpu(hctx);
 809                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 810         }
 811 }
 812
 813 void blk_mq_run_queues(struct request_queue *q, bool async)
 814 {
 815         struct blk_mq_hw_ctx *hctx;
 816         int i;
 817
 818         queue_for_each_hw_ctx(q, hctx, i) {
 819                 if ((!blk_mq_hctx_has_pending(hctx) &&
 820                     list_empty_careful(&hctx->dispatch)) ||
 821                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 822                         continue;
 823
 824                 preempt_disable();
 825                 blk_mq_run_hw_queue(hctx, async);
 826                 preempt_enable();
 827         }
 828 }
 829 EXPORT_SYMBOL(blk_mq_run_queues);
 830
 831 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 832 {
 833         cancel_delayed_work(&hctx->run_work);
 834         cancel_delayed_work(&hctx->delay_work);
 835         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 836 }
 837 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 838
 839 void blk_mq_stop_hw_queues(struct request_queue *q)
 840 {
 841         struct blk_mq_hw_ctx *hctx;
 842         int i;
 843
 844         queue_for_each_hw_ctx(q, hctx, i)
 845                 blk_mq_stop_hw_queue(hctx);
 846 }
 847 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 848
 849 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 850 {
 851         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 852
 853         preempt_disable();
 854         blk_mq_run_hw_queue(hctx, false);
 855         preempt_enable();
 856 }
 857 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 858
 859 void blk_mq_start_hw_queues(struct request_queue *q)
 860 {
 861         struct blk_mq_hw_ctx *hctx;
 862         int i;
 863
 864         queue_for_each_hw_ctx(q, hctx, i)
 865                 blk_mq_start_hw_queue(hctx);
 866 }
 867 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 868
 869
 870 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 871 {
 872         struct blk_mq_hw_ctx *hctx;
 873         int i;
 874
 875         queue_for_each_hw_ctx(q, hctx, i) {
 876                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 877                         continue;
 878
 879                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 880                 preempt_disable();
 881                 blk_mq_run_hw_queue(hctx, async);
 882                 preempt_enable();
 883         }
 884 }
 885 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 886
 887 static void blk_mq_run_work_fn(struct work_struct *work)
 888 {
 889         struct blk_mq_hw_ctx *hctx;
 890
 891         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 892
 893         __blk_mq_run_hw_queue(hctx);
 894 }
 895
 896 static void blk_mq_delay_work_fn(struct work_struct *work)
 897 {
 898         struct blk_mq_hw_ctx *hctx;
 899
 900         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 901
 902         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 903                 __blk_mq_run_hw_queue(hctx);
 904 }
 905
 906 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 907 {
 908         unsigned long tmo = msecs_to_jiffies(msecs);
 909
 910         if (hctx->queue->nr_hw_queues == 1)
 911                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 912         else {
 913                 unsigned int cpu;
 914
 915                 cpu = blk_mq_hctx_next_cpu(hctx);
 916                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 917         }
 918 }
 919 EXPORT_SYMBOL(blk_mq_delay_queue);
 920
 921 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 922                                     struct request *rq, bool at_head)
 923 {
 924         struct blk_mq_ctx *ctx = rq->mq_ctx;
 925
 926         trace_block_rq_insert(hctx->queue, rq);
 927
 928         if (at_head)
 929                 list_add(&rq->queuelist, &ctx->rq_list);
 930         else
 931                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 932
 933         blk_mq_hctx_mark_pending(hctx, ctx);
 934 }
 935
 936 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 937                 bool async)
 938 {
 939         struct request_queue *q = rq->q;
 940         struct blk_mq_hw_ctx *hctx;
 941         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 942
 943         current_ctx = blk_mq_get_ctx(q);
 944         if (!cpu_online(ctx->cpu))
 945                 rq->mq_ctx = ctx = current_ctx;
 946
 947         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 948
 949         spin_lock(&ctx->lock);
 950         __blk_mq_insert_request(hctx, rq, at_head);
 951         spin_unlock(&ctx->lock);
 952
 953         if (run_queue)
 954                 blk_mq_run_hw_queue(hctx, async);
 955
 956         blk_mq_put_ctx(current_ctx);
 957 }
 958
 959 static void blk_mq_insert_requests(struct request_queue *q,
 960                                      struct blk_mq_ctx *ctx,
 961                                      struct list_head *list,
 962                                      int depth,
 963                                      bool from_schedule)
 964
 965 {
 966         struct blk_mq_hw_ctx *hctx;
 967         struct blk_mq_ctx *current_ctx;
 968
 969         trace_block_unplug(q, depth, !from_schedule);
 970
 971         current_ctx = blk_mq_get_ctx(q);
 972
 973         if (!cpu_online(ctx->cpu))
 974                 ctx = current_ctx;
 975         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 976
 977         /*
 978          * preemption doesn't flush plug list, so it's possible ctx->cpu is
 979          * offline now
 980          */
 981         spin_lock(&ctx->lock);
 982         while (!list_empty(list)) {
 983                 struct request *rq;
 984
 985                 rq = list_first_entry(list, struct request, queuelist);
 986                 list_del_init(&rq->queuelist);
 987                 rq->mq_ctx = ctx;
 988                 __blk_mq_insert_request(hctx, rq, false);
 989         }
 990         spin_unlock(&ctx->lock);
 991
 992         blk_mq_run_hw_queue(hctx, from_schedule);
 993         blk_mq_put_ctx(current_ctx);
 994 }
 995
 996 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
 997 {
 998         struct request *rqa = container_of(a, struct request, queuelist);
 999         struct request *rqb = container_of(b, struct request, queuelist);
1000
1001         return !(rqa->mq_ctx < rqb->mq_ctx ||
1002                  (rqa->mq_ctx == rqb->mq_ctx &&
1003                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1004 }
1005
1006 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1007 {
1008         struct blk_mq_ctx *this_ctx;
1009         struct request_queue *this_q;
1010         struct request *rq;
1011         LIST_HEAD(list);
1012         LIST_HEAD(ctx_list);
1013         unsigned int depth;
1014
1015         list_splice_init(&plug->mq_list, &list);
1016
1017         list_sort(NULL, &list, plug_ctx_cmp);
1018
1019         this_q = NULL;
1020         this_ctx = NULL;
1021         depth = 0;
1022
1023         while (!list_empty(&list)) {
1024                 rq = list_entry_rq(list.next);
1025                 list_del_init(&rq->queuelist);
1026                 BUG_ON(!rq->q);
1027                 if (rq->mq_ctx != this_ctx) {
1028                         if (this_ctx) {
1029                                 blk_mq_insert_requests(this_q, this_ctx,
1030                                                         &ctx_list, depth,
1031                                                         from_schedule);
1032                         }
1033
1034                         this_ctx = rq->mq_ctx;
1035                         this_q = rq->q;
1036                         depth = 0;
1037                 }
1038
1039                 depth++;
1040                 list_add_tail(&rq->queuelist, &ctx_list);
1041         }
1042
1043         /*
1044          * If 'this_ctx' is set, we know we have entries to complete
1045          * on 'ctx_list'. Do those.
1046          */
1047         if (this_ctx) {
1048                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1049                                        from_schedule);
1050         }
1051 }
1052
1053 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1054 {
1055         init_request_from_bio(rq, bio);
1056
1057         if (blk_do_io_stat(rq))
1058                 blk_account_io_start(rq, 1);
1059 }
1060
1061 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1062 {
1063         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1064                 !blk_queue_nomerges(hctx->queue);
1065 }
1066
1067 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1068                                          struct blk_mq_ctx *ctx,
1069                                          struct request *rq, struct bio *bio)
1070 {
1071         if (!hctx_allow_merges(hctx)) {
1072                 blk_mq_bio_to_request(rq, bio);
1073                 spin_lock(&ctx->lock);
1074 insert_rq:
1075                 __blk_mq_insert_request(hctx, rq, false);
1076                 spin_unlock(&ctx->lock);
1077                 return false;
1078         } else {
1079                 struct request_queue *q = hctx->queue;
1080
1081                 spin_lock(&ctx->lock);
1082                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1083                         blk_mq_bio_to_request(rq, bio);
1084                         goto insert_rq;
1085                 }
1086
1087                 spin_unlock(&ctx->lock);
1088                 __blk_mq_free_request(hctx, ctx, rq);
1089                 return true;
1090         }
1091 }
1092
1093 struct blk_map_ctx {
1094         struct blk_mq_hw_ctx *hctx;
1095         struct blk_mq_ctx *ctx;
1096 };
1097
1098 static struct request *blk_mq_map_request(struct request_queue *q,
1099                                           struct bio *bio,
1100                                           struct blk_map_ctx *data)
1101 {
1102         struct blk_mq_hw_ctx *hctx;
1103         struct blk_mq_ctx *ctx;
1104         struct request *rq;
1105         int rw = bio_data_dir(bio);
1106         struct blk_mq_alloc_data alloc_data;
1107
1108         if (unlikely(blk_mq_queue_enter(q))) {
1109                 bio_endio(bio, -EIO);
1110                 return NULL;
1111         }
1112
1113         ctx = blk_mq_get_ctx(q);
1114         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1115
1116         if (rw_is_sync(bio->bi_rw))
1117                 rw |= REQ_SYNC;
1118
1119         trace_block_getrq(q, bio, rw);
1120         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1121                         hctx);
1122         rq = __blk_mq_alloc_request(&alloc_data, rw);
1123         if (unlikely(!rq)) {
1124                 __blk_mq_run_hw_queue(hctx);
1125                 blk_mq_put_ctx(ctx);
1126                 trace_block_sleeprq(q, bio, rw);
1127
1128                 ctx = blk_mq_get_ctx(q);
1129                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1130                 blk_mq_set_alloc_data(&alloc_data, q,
1131                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1132                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1133                 ctx = alloc_data.ctx;
1134                 hctx = alloc_data.hctx;
1135         }
1136
1137         hctx->queued++;
1138         data->hctx = hctx;
1139         data->ctx = ctx;
1140         return rq;
1141 }
1142
1143 /*
1144  * Multiple hardware queue variant. This will not use per-process plugs,
1145  * but will attempt to bypass the hctx queueing if we can go straight to
1146  * hardware for SYNC IO.
1147  */
1148 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1149 {
1150         const int is_sync = rw_is_sync(bio->bi_rw);
1151         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1152         struct blk_map_ctx data;
1153         struct request *rq;
1154
1155         blk_queue_bounce(q, &bio);
1156
1157         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1158                 bio_endio(bio, -EIO);
1159                 return;
1160         }
1161
1162         rq = blk_mq_map_request(q, bio, &data);
1163         if (unlikely(!rq))
1164                 return;
1165
1166         if (unlikely(is_flush_fua)) {
1167                 blk_mq_bio_to_request(rq, bio);
1168                 blk_insert_flush(rq);
1169                 goto run_queue;
1170         }
1171
1172         if (is_sync) {
1173                 int ret;
1174
1175                 blk_mq_bio_to_request(rq, bio);
1176
1177                 /*
1178                  * For OK queue, we are done. For error, kill it. Any other
1179                  * error (busy), just add it to our list as we previously
1180                  * would have done
1181                  */
1182                 ret = q->mq_ops->queue_rq(data.hctx, rq, true);
1183                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1184                         goto done;
1185                 else {
1186                         __blk_mq_requeue_request(rq);
1187
1188                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1189                                 rq->errors = -EIO;
1190                                 blk_mq_end_request(rq, rq->errors);
1191                                 goto done;
1192                         }
1193                 }
1194         }
1195
1196         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1197                 /*
1198                  * For a SYNC request, send it to the hardware immediately. For
1199                  * an ASYNC request, just ensure that we run it later on. The
1200                  * latter allows for merging opportunities and more efficient
1201                  * dispatching.
1202                  */
1203 run_queue:
1204                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1205         }
1206 done:
1207         blk_mq_put_ctx(data.ctx);
1208 }
1209
1210 /*
1211  * Single hardware queue variant. This will attempt to use any per-process
1212  * plug for merging and IO deferral.
1213  */
1214 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1215 {
1216         const int is_sync = rw_is_sync(bio->bi_rw);
1217         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1218         unsigned int use_plug, request_count = 0;
1219         struct blk_map_ctx data;
1220         struct request *rq;
1221
1222         /*
1223          * If we have multiple hardware queues, just go directly to
1224          * one of those for sync IO.
1225          */
1226         use_plug = !is_flush_fua && !is_sync;
1227
1228         blk_queue_bounce(q, &bio);
1229
1230         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1231                 bio_endio(bio, -EIO);
1232                 return;
1233         }
1234
1235         if (use_plug && !blk_queue_nomerges(q) &&
1236             blk_attempt_plug_merge(q, bio, &request_count))
1237                 return;
1238
1239         rq = blk_mq_map_request(q, bio, &data);
1240         if (unlikely(!rq))
1241                 return;
1242
1243         if (unlikely(is_flush_fua)) {
1244                 blk_mq_bio_to_request(rq, bio);
1245                 blk_insert_flush(rq);
1246                 goto run_queue;
1247         }
1248
1249         /*
1250          * A task plug currently exists. Since this is completely lockless,
1251          * utilize that to temporarily store requests until the task is
1252          * either done or scheduled away.
1253          */
1254         if (use_plug) {
1255                 struct blk_plug *plug = current->plug;
1256
1257                 if (plug) {
1258                         blk_mq_bio_to_request(rq, bio);
1259                         if (list_empty(&plug->mq_list))
1260                                 trace_block_plug(q);
1261                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1262                                 blk_flush_plug_list(plug, false);
1263                                 trace_block_plug(q);
1264                         }
1265                         list_add_tail(&rq->queuelist, &plug->mq_list);
1266                         blk_mq_put_ctx(data.ctx);
1267                         return;
1268                 }
1269         }
1270
1271         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1272                 /*
1273                  * For a SYNC request, send it to the hardware immediately. For
1274                  * an ASYNC request, just ensure that we run it later on. The
1275                  * latter allows for merging opportunities and more efficient
1276                  * dispatching.
1277                  */
1278 run_queue:
1279                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1280         }
1281
1282         blk_mq_put_ctx(data.ctx);
1283 }
1284
1285 /*
1286  * Default mapping to a software queue, since we use one per CPU.
1287  */
1288 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1289 {
1290         return q->queue_hw_ctx[q->mq_map[cpu]];
1291 }
1292 EXPORT_SYMBOL(blk_mq_map_queue);
1293
1294 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1295                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1296 {
1297         struct page *page;
1298
1299         if (tags->rqs && set->ops->exit_request) {
1300                 int i;
1301
1302                 for (i = 0; i < tags->nr_tags; i++) {
1303                         if (!tags->rqs[i])
1304                                 continue;
1305                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1306                                                 hctx_idx, i);
1307                         tags->rqs[i] = NULL;
1308                 }
1309         }
1310
1311         while (!list_empty(&tags->page_list)) {
1312                 page = list_first_entry(&tags->page_list, struct page, lru);
1313                 list_del_init(&page->lru);
1314                 __free_pages(page, page->private);
1315         }
1316
1317         kfree(tags->rqs);
1318
1319         blk_mq_free_tags(tags);
1320 }
1321
1322 static size_t order_to_size(unsigned int order)
1323 {
1324         return (size_t)PAGE_SIZE << order;
1325 }
1326
1327 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1328                 unsigned int hctx_idx)
1329 {
1330         struct blk_mq_tags *tags;
1331         unsigned int i, j, entries_per_page, max_order = 4;
1332         size_t rq_size, left;
1333
1334         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1335                                 set->numa_node);
1336         if (!tags)
1337                 return NULL;
1338
1339         INIT_LIST_HEAD(&tags->page_list);
1340
1341         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1342                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1343                                  set->numa_node);
1344         if (!tags->rqs) {
1345                 blk_mq_free_tags(tags);
1346                 return NULL;
1347         }
1348
1349         /*
1350          * rq_size is the size of the request plus driver payload, rounded
1351          * to the cacheline size
1352          */
1353         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1354                                 cache_line_size());
1355         left = rq_size * set->queue_depth;
1356
1357         for (i = 0; i < set->queue_depth; ) {
1358                 int this_order = max_order;
1359                 struct page *page;
1360                 int to_do;
1361                 void *p;
1362
1363                 while (left < order_to_size(this_order - 1) && this_order)
1364                         this_order--;
1365
1366                 do {
1367                         page = alloc_pages_node(set->numa_node,
1368                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1369                                 this_order);
1370                         if (page)
1371                                 break;
1372                         if (!this_order--)
1373                                 break;
1374                         if (order_to_size(this_order) < rq_size)
1375                                 break;
1376                 } while (1);
1377
1378                 if (!page)
1379                         goto fail;
1380
1381                 page->private = this_order;
1382                 list_add_tail(&page->lru, &tags->page_list);
1383
1384                 p = page_address(page);
1385                 entries_per_page = order_to_size(this_order) / rq_size;
1386                 to_do = min(entries_per_page, set->queue_depth - i);
1387                 left -= to_do * rq_size;
1388                 for (j = 0; j < to_do; j++) {
1389                         tags->rqs[i] = p;
1390                         tags->rqs[i]->atomic_flags = 0;
1391                         tags->rqs[i]->cmd_flags = 0;
1392                         if (set->ops->init_request) {
1393                                 if (set->ops->init_request(set->driver_data,
1394                                                 tags->rqs[i], hctx_idx, i,
1395                                                 set->numa_node)) {
1396                                         tags->rqs[i] = NULL;
1397                                         goto fail;
1398                                 }
1399                         }
1400
1401                         p += rq_size;
1402                         i++;
1403                 }
1404         }
1405
1406         return tags;
1407
1408 fail:
1409         blk_mq_free_rq_map(set, tags, hctx_idx);
1410         return NULL;
1411 }
1412
1413 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1414 {
1415         kfree(bitmap->map);
1416 }
1417
1418 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1419 {
1420         unsigned int bpw = 8, total, num_maps, i;
1421
1422         bitmap->bits_per_word = bpw;
1423
1424         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1425         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1426                                         GFP_KERNEL, node);
1427         if (!bitmap->map)
1428                 return -ENOMEM;
1429
1430         bitmap->map_size = num_maps;
1431
1432         total = nr_cpu_ids;
1433         for (i = 0; i < num_maps; i++) {
1434                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1435                 total -= bitmap->map[i].depth;
1436         }
1437
1438         return 0;
1439 }
1440
1441 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1442 {
1443         struct request_queue *q = hctx->queue;
1444         struct blk_mq_ctx *ctx;
1445         LIST_HEAD(tmp);
1446
1447         /*
1448          * Move ctx entries to new CPU, if this one is going away.
1449          */
1450         ctx = __blk_mq_get_ctx(q, cpu);
1451
1452         spin_lock(&ctx->lock);
1453         if (!list_empty(&ctx->rq_list)) {
1454                 list_splice_init(&ctx->rq_list, &tmp);
1455                 blk_mq_hctx_clear_pending(hctx, ctx);
1456         }
1457         spin_unlock(&ctx->lock);
1458
1459         if (list_empty(&tmp))
1460                 return NOTIFY_OK;
1461
1462         ctx = blk_mq_get_ctx(q);
1463         spin_lock(&ctx->lock);
1464
1465         while (!list_empty(&tmp)) {
1466                 struct request *rq;
1467
1468                 rq = list_first_entry(&tmp, struct request, queuelist);
1469                 rq->mq_ctx = ctx;
1470                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1471         }
1472
1473         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1474         blk_mq_hctx_mark_pending(hctx, ctx);
1475
1476         spin_unlock(&ctx->lock);
1477
1478         blk_mq_run_hw_queue(hctx, true);
1479         blk_mq_put_ctx(ctx);
1480         return NOTIFY_OK;
1481 }
1482
1483 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1484 {
1485         struct request_queue *q = hctx->queue;
1486         struct blk_mq_tag_set *set = q->tag_set;
1487
1488         if (set->tags[hctx->queue_num])
1489                 return NOTIFY_OK;
1490
1491         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1492         if (!set->tags[hctx->queue_num])
1493                 return NOTIFY_STOP;
1494
1495         hctx->tags = set->tags[hctx->queue_num];
1496         return NOTIFY_OK;
1497 }
1498
1499 static int blk_mq_hctx_notify(void *data, unsigned long action,
1500                               unsigned int cpu)
1501 {
1502         struct blk_mq_hw_ctx *hctx = data;
1503
1504         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1505                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1506         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1507                 return blk_mq_hctx_cpu_online(hctx, cpu);
1508
1509         return NOTIFY_OK;
1510 }
1511
1512 static void blk_mq_exit_hw_queues(struct request_queue *q,
1513                 struct blk_mq_tag_set *set, int nr_queue)
1514 {
1515         struct blk_mq_hw_ctx *hctx;
1516         unsigned int i;
1517
1518         queue_for_each_hw_ctx(q, hctx, i) {
1519                 if (i == nr_queue)
1520                         break;
1521
1522                 blk_mq_tag_idle(hctx);
1523
1524                 if (set->ops->exit_hctx)
1525                         set->ops->exit_hctx(hctx, i);
1526
1527                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1528                 kfree(hctx->ctxs);
1529                 blk_mq_free_bitmap(&hctx->ctx_map);
1530         }
1531
1532 }
1533
1534 static void blk_mq_free_hw_queues(struct request_queue *q,
1535                 struct blk_mq_tag_set *set)
1536 {
1537         struct blk_mq_hw_ctx *hctx;
1538         unsigned int i;
1539
1540         queue_for_each_hw_ctx(q, hctx, i) {
1541                 free_cpumask_var(hctx->cpumask);
1542                 kfree(hctx);
1543         }
1544 }
1545
1546 static int blk_mq_init_hw_queues(struct request_queue *q,
1547                 struct blk_mq_tag_set *set)
1548 {
1549         struct blk_mq_hw_ctx *hctx;
1550         unsigned int i;
1551
1552         /*
1553          * Initialize hardware queues
1554          */
1555         queue_for_each_hw_ctx(q, hctx, i) {
1556                 int node;
1557
1558                 node = hctx->numa_node;
1559                 if (node == NUMA_NO_NODE)
1560                         node = hctx->numa_node = set->numa_node;
1561
1562                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1563                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1564                 spin_lock_init(&hctx->lock);
1565                 INIT_LIST_HEAD(&hctx->dispatch);
1566                 hctx->queue = q;
1567                 hctx->queue_num = i;
1568                 hctx->flags = set->flags;
1569                 hctx->cmd_size = set->cmd_size;
1570
1571                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1572                                                 blk_mq_hctx_notify, hctx);
1573                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1574
1575                 hctx->tags = set->tags[i];
1576
1577                 /*
1578                  * Allocate space for all possible cpus to avoid allocation at
1579                  * runtime
1580                  */
1581                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1582                                                 GFP_KERNEL, node);
1583                 if (!hctx->ctxs)
1584                         break;
1585
1586                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1587                         break;
1588
1589                 hctx->nr_ctx = 0;
1590
1591                 if (set->ops->init_hctx &&
1592                     set->ops->init_hctx(hctx, set->driver_data, i))
1593                         break;
1594         }
1595
1596         if (i == q->nr_hw_queues)
1597                 return 0;
1598
1599         /*
1600          * Init failed
1601          */
1602         blk_mq_exit_hw_queues(q, set, i);
1603
1604         return 1;
1605 }
1606
1607 static void blk_mq_init_cpu_queues(struct request_queue *q,
1608                                    unsigned int nr_hw_queues)
1609 {
1610         unsigned int i;
1611
1612         for_each_possible_cpu(i) {
1613                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1614                 struct blk_mq_hw_ctx *hctx;
1615
1616                 memset(__ctx, 0, sizeof(*__ctx));
1617                 __ctx->cpu = i;
1618                 spin_lock_init(&__ctx->lock);
1619                 INIT_LIST_HEAD(&__ctx->rq_list);
1620                 __ctx->queue = q;
1621
1622                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1623                 if (!cpu_online(i))
1624                         continue;
1625
1626                 hctx = q->mq_ops->map_queue(q, i);
1627                 cpumask_set_cpu(i, hctx->cpumask);
1628                 hctx->nr_ctx++;
1629
1630                 /*
1631                  * Set local node, IFF we have more than one hw queue. If
1632                  * not, we remain on the home node of the device
1633                  */
1634                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1635                         hctx->numa_node = cpu_to_node(i);
1636         }
1637 }
1638
1639 static void blk_mq_map_swqueue(struct request_queue *q)
1640 {
1641         unsigned int i;
1642         struct blk_mq_hw_ctx *hctx;
1643         struct blk_mq_ctx *ctx;
1644
1645         queue_for_each_hw_ctx(q, hctx, i) {
1646                 cpumask_clear(hctx->cpumask);
1647                 hctx->nr_ctx = 0;
1648         }
1649
1650         /*
1651          * Map software to hardware queues
1652          */
1653         queue_for_each_ctx(q, ctx, i) {
1654                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1655                 if (!cpu_online(i))
1656                         continue;
1657
1658                 hctx = q->mq_ops->map_queue(q, i);
1659                 cpumask_set_cpu(i, hctx->cpumask);
1660                 ctx->index_hw = hctx->nr_ctx;
1661                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1662         }
1663
1664         queue_for_each_hw_ctx(q, hctx, i) {
1665                 /*
1666                  * If no software queues are mapped to this hardware queue,
1667                  * disable it and free the request entries.
1668                  */
1669                 if (!hctx->nr_ctx) {
1670                         struct blk_mq_tag_set *set = q->tag_set;
1671
1672                         if (set->tags[i]) {
1673                                 blk_mq_free_rq_map(set, set->tags[i], i);
1674                                 set->tags[i] = NULL;
1675                                 hctx->tags = NULL;
1676                         }
1677                         continue;
1678                 }
1679
1680                 /*
1681                  * Initialize batch roundrobin counts
1682                  */
1683                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1684                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1685         }
1686 }
1687
1688 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1689 {
1690         struct blk_mq_hw_ctx *hctx;
1691         struct request_queue *q;
1692         bool shared;
1693         int i;
1694
1695         if (set->tag_list.next == set->tag_list.prev)
1696                 shared = false;
1697         else
1698                 shared = true;
1699
1700         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1701                 blk_mq_freeze_queue(q);
1702
1703                 queue_for_each_hw_ctx(q, hctx, i) {
1704                         if (shared)
1705                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1706                         else
1707                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1708                 }
1709                 blk_mq_unfreeze_queue(q);
1710         }
1711 }
1712
1713 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1714 {
1715         struct blk_mq_tag_set *set = q->tag_set;
1716
1717         mutex_lock(&set->tag_list_lock);
1718         list_del_init(&q->tag_set_list);
1719         blk_mq_update_tag_set_depth(set);
1720         mutex_unlock(&set->tag_list_lock);
1721 }
1722
1723 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1724                                      struct request_queue *q)
1725 {
1726         q->tag_set = set;
1727
1728         mutex_lock(&set->tag_list_lock);
1729         list_add_tail(&q->tag_set_list, &set->tag_list);
1730         blk_mq_update_tag_set_depth(set);
1731         mutex_unlock(&set->tag_list_lock);
1732 }
1733
1734 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1735 {
1736         struct blk_mq_hw_ctx **hctxs;
1737         struct blk_mq_ctx __percpu *ctx;
1738         struct request_queue *q;
1739         unsigned int *map;
1740         int i;
1741
1742         ctx = alloc_percpu(struct blk_mq_ctx);
1743         if (!ctx)
1744                 return ERR_PTR(-ENOMEM);
1745
1746         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1747                         set->numa_node);
1748
1749         if (!hctxs)
1750                 goto err_percpu;
1751
1752         map = blk_mq_make_queue_map(set);
1753         if (!map)
1754                 goto err_map;
1755
1756         for (i = 0; i < set->nr_hw_queues; i++) {
1757                 int node = blk_mq_hw_queue_to_node(map, i);
1758
1759                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1760                                         GFP_KERNEL, node);
1761                 if (!hctxs[i])
1762                         goto err_hctxs;
1763
1764                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1765                         goto err_hctxs;
1766
1767                 atomic_set(&hctxs[i]->nr_active, 0);
1768                 hctxs[i]->numa_node = node;
1769                 hctxs[i]->queue_num = i;
1770         }
1771
1772         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1773         if (!q)
1774                 goto err_hctxs;
1775
1776         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1777                 goto err_map;
1778
1779         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1780         blk_queue_rq_timeout(q, 30000);
1781
1782         q->nr_queues = nr_cpu_ids;
1783         q->nr_hw_queues = set->nr_hw_queues;
1784         q->mq_map = map;
1785
1786         q->queue_ctx = ctx;
1787         q->queue_hw_ctx = hctxs;
1788
1789         q->mq_ops = set->ops;
1790         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1791
1792         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1793                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1794
1795         q->sg_reserved_size = INT_MAX;
1796
1797         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1798         INIT_LIST_HEAD(&q->requeue_list);
1799         spin_lock_init(&q->requeue_lock);
1800
1801         if (q->nr_hw_queues > 1)
1802                 blk_queue_make_request(q, blk_mq_make_request);
1803         else
1804                 blk_queue_make_request(q, blk_sq_make_request);
1805
1806         if (set->timeout)
1807                 blk_queue_rq_timeout(q, set->timeout);
1808
1809         /*
1810          * Do this after blk_queue_make_request() overrides it...
1811          */
1812         q->nr_requests = set->queue_depth;
1813
1814         if (set->ops->complete)
1815                 blk_queue_softirq_done(q, set->ops->complete);
1816
1817         blk_mq_init_flush(q);
1818         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1819
1820         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1821                                 set->cmd_size, cache_line_size()),
1822                                 GFP_KERNEL);
1823         if (!q->flush_rq)
1824                 goto err_hw;
1825
1826         if (blk_mq_init_hw_queues(q, set))
1827                 goto err_flush_rq;
1828
1829         mutex_lock(&all_q_mutex);
1830         list_add_tail(&q->all_q_node, &all_q_list);
1831         mutex_unlock(&all_q_mutex);
1832
1833         blk_mq_add_queue_tag_set(set, q);
1834
1835         blk_mq_map_swqueue(q);
1836
1837         return q;
1838
1839 err_flush_rq:
1840         kfree(q->flush_rq);
1841 err_hw:
1842         blk_cleanup_queue(q);
1843 err_hctxs:
1844         kfree(map);
1845         for (i = 0; i < set->nr_hw_queues; i++) {
1846                 if (!hctxs[i])
1847                         break;
1848                 free_cpumask_var(hctxs[i]->cpumask);
1849                 kfree(hctxs[i]);
1850         }
1851 err_map:
1852         kfree(hctxs);
1853 err_percpu:
1854         free_percpu(ctx);
1855         return ERR_PTR(-ENOMEM);
1856 }
1857 EXPORT_SYMBOL(blk_mq_init_queue);
1858
1859 void blk_mq_free_queue(struct request_queue *q)
1860 {
1861         struct blk_mq_tag_set   *set = q->tag_set;
1862
1863         blk_mq_del_queue_tag_set(q);
1864
1865         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1866         blk_mq_free_hw_queues(q, set);
1867
1868         percpu_ref_exit(&q->mq_usage_counter);
1869
1870         free_percpu(q->queue_ctx);
1871         kfree(q->queue_hw_ctx);
1872         kfree(q->mq_map);
1873
1874         q->queue_ctx = NULL;
1875         q->queue_hw_ctx = NULL;
1876         q->mq_map = NULL;
1877
1878         mutex_lock(&all_q_mutex);
1879         list_del_init(&q->all_q_node);
1880         mutex_unlock(&all_q_mutex);
1881 }
1882
1883 /* Basically redo blk_mq_init_queue with queue frozen */
1884 static void blk_mq_queue_reinit(struct request_queue *q)
1885 {
1886         blk_mq_freeze_queue(q);
1887
1888         blk_mq_sysfs_unregister(q);
1889
1890         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1891
1892         /*
1893          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1894          * we should change hctx numa_node according to new topology (this
1895          * involves free and re-allocate memory, worthy doing?)
1896          */
1897
1898         blk_mq_map_swqueue(q);
1899
1900         blk_mq_sysfs_register(q);
1901
1902         blk_mq_unfreeze_queue(q);
1903 }
1904
1905 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1906                                       unsigned long action, void *hcpu)
1907 {
1908         struct request_queue *q;
1909
1910         /*
1911          * Before new mappings are established, hotadded cpu might already
1912          * start handling requests. This doesn't break anything as we map
1913          * offline CPUs to first hardware queue. We will re-init the queue
1914          * below to get optimal settings.
1915          */
1916         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1917             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1918                 return NOTIFY_OK;
1919
1920         mutex_lock(&all_q_mutex);
1921         list_for_each_entry(q, &all_q_list, all_q_node)
1922                 blk_mq_queue_reinit(q);
1923         mutex_unlock(&all_q_mutex);
1924         return NOTIFY_OK;
1925 }
1926
1927 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1928 {
1929         int i;
1930
1931         for (i = 0; i < set->nr_hw_queues; i++) {
1932                 set->tags[i] = blk_mq_init_rq_map(set, i);
1933                 if (!set->tags[i])
1934                         goto out_unwind;
1935         }
1936
1937         return 0;
1938
1939 out_unwind:
1940         while (--i >= 0)
1941                 blk_mq_free_rq_map(set, set->tags[i], i);
1942
1943         return -ENOMEM;
1944 }
1945
1946 /*
1947  * Allocate the request maps associated with this tag_set. Note that this
1948  * may reduce the depth asked for, if memory is tight. set->queue_depth
1949  * will be updated to reflect the allocated depth.
1950  */
1951 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1952 {
1953         unsigned int depth;
1954         int err;
1955
1956         depth = set->queue_depth;
1957         do {
1958                 err = __blk_mq_alloc_rq_maps(set);
1959                 if (!err)
1960                         break;
1961
1962                 set->queue_depth >>= 1;
1963                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1964                         err = -ENOMEM;
1965                         break;
1966                 }
1967         } while (set->queue_depth);
1968
1969         if (!set->queue_depth || err) {
1970                 pr_err("blk-mq: failed to allocate request map\n");
1971                 return -ENOMEM;
1972         }
1973
1974         if (depth != set->queue_depth)
1975                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
1976                                                 depth, set->queue_depth);
1977
1978         return 0;
1979 }
1980
1981 /*
1982  * Alloc a tag set to be associated with one or more request queues.
1983  * May fail with EINVAL for various error conditions. May adjust the
1984  * requested depth down, if if it too large. In that case, the set
1985  * value will be stored in set->queue_depth.
1986  */
1987 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1988 {
1989         if (!set->nr_hw_queues)
1990                 return -EINVAL;
1991         if (!set->queue_depth)
1992                 return -EINVAL;
1993         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1994                 return -EINVAL;
1995
1996         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1997                 return -EINVAL;
1998
1999         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2000                 pr_info("blk-mq: reduced tag depth to %u\n",
2001                         BLK_MQ_MAX_DEPTH);
2002                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2003         }
2004
2005         set->tags = kmalloc_node(set->nr_hw_queues *
2006                                  sizeof(struct blk_mq_tags *),
2007                                  GFP_KERNEL, set->numa_node);
2008         if (!set->tags)
2009                 return -ENOMEM;
2010
2011         if (blk_mq_alloc_rq_maps(set))
2012                 goto enomem;
2013
2014         mutex_init(&set->tag_list_lock);
2015         INIT_LIST_HEAD(&set->tag_list);
2016
2017         return 0;
2018 enomem:
2019         kfree(set->tags);
2020         set->tags = NULL;
2021         return -ENOMEM;
2022 }
2023 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2024
2025 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2026 {
2027         int i;
2028
2029         for (i = 0; i < set->nr_hw_queues; i++) {
2030                 if (set->tags[i])
2031                         blk_mq_free_rq_map(set, set->tags[i], i);
2032         }
2033
2034         kfree(set->tags);
2035         set->tags = NULL;
2036 }
2037 EXPORT_SYMBOL(blk_mq_free_tag_set);
2038
2039 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2040 {
2041         struct blk_mq_tag_set *set = q->tag_set;
2042         struct blk_mq_hw_ctx *hctx;
2043         int i, ret;
2044
2045         if (!set || nr > set->queue_depth)
2046                 return -EINVAL;
2047
2048         ret = 0;
2049         queue_for_each_hw_ctx(q, hctx, i) {
2050                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2051                 if (ret)
2052                         break;
2053         }
2054
2055         if (!ret)
2056                 q->nr_requests = nr;
2057
2058         return ret;
2059 }
2060
2061 void blk_mq_disable_hotplug(void)
2062 {
2063         mutex_lock(&all_q_mutex);
2064 }
2065
2066 void blk_mq_enable_hotplug(void)
2067 {
2068         mutex_unlock(&all_q_mutex);
2069 }
2070
2071 static int __init blk_mq_init(void)
2072 {
2073         blk_mq_cpu_init();
2074
2075         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2076
2077         return 0;
2078 }
2079 subsys_initcall(blk_mq_init);